ChatGPT Work tests the limits of office permissions, OpenAI's bet to move programming agents to all employees

OpenAI aims to move independent agent technologies from developer environments to traditional offices by launching its product ChatGPT Work, available as part of the basic subscription plan for $20 per month. This product is a modified version of the programming tool Codex, intended to enable accountants, investors, and operations managers to integrate language models into their daily workflows by controlling email, Slack platforms, and applications such as Notion and Figma, and to execute multi-step tasks autonomously.

Shifting from merely answering questions to autonomously managing complex projects is the biggest commercial bet for labs today.Internal company data show a large gap between internal and external use; a study funded by OpenAI revealed that 98 percent of its employees use the Codex tool, while adoption among enterprise subscribers does not exceed 17 percent and is below 1 percent among individuals. While the total number of desktop app users is around 20 million, the traditional web interface has over a billion users, reflecting the difficulty of convincing the average user to delegate their tasks to a software tool.

The company faces an engineering challenge in designing what is called a software harness, the layer that defines the tools and permissions granted to the model. Unlike programmers who are satisfied with command-line interfaces, scaling agents requires interactive interfaces that reduce friction and make benefits easier to discover. This effort runs into complex access permissions, as tools often require granting full read-write rights to cloud files instead of limited read-only permissions, alongside the difficulty of measuring the quality of office-work outputs such as strategies and presentations compared with testable code, which the company attempts to assess through the GDPval benchmark covering 44 occupations and hundreds of cognitive tests.

Running agents for longer periods consumes larger numbers of tokens, which forms the financial return pillar used to justify massive investments in training and data centers.Competition intensifies in this space with specialized tools such as Harvey for legal affairs and Klai for sales, as well as competing products like Claude Cowork, all of which aim to capture added value by integrating models into enterprise data flows.

For business sectors and institutions in the Gulf, Egypt, and the Levant, this shift forces a reassessment of information security policies and access governance. Moving from using AI as a search or text-draft interface to granting it direct integration with company records and internal communication channels means technology departments will face a decisive choice between strict security that hampers agent efficiency and broader permissions that risk exposing models to sensitive financial and administrative data. It also requires professionals in investment, accounting, and operations to move from composing simple textual commands to skills in managing and auditing automated workflow streams and ensuring the reliability of results executed by systems on their behalf.

The success of this bet remains contingent on the models' ability to demonstrate viability in handling daily data chaos without making catastrophic errors, making the balance between granting permissions and preserving privacy the true benchmark for adopting these tools beyond narrow technical teams.